Intelligent Vehicle Adaptive Kalman Filtering Multi-Object Dynamic Tracking Method With Fusion Of Appearance Features

Multi-object tracking technology is an important component of the environmental perception technology in vehicle intelligent technology.It can provide intelligent vehicles with recognition and tracking of multiple objects such as pedestrians,vehicles,and traffic signs,making driving safer and more reliable.However,in dynamic road scenarios,the camera’s motion can lead to inaccurate predictions of the motion model,and the recognition ability of the object features extracted by the appearance feature extraction model is also insufficient.In order to address these issues,this paper proposes a camera motion compensation method that combines LK optical flow with Fast feature points,as well as a multi-object tracking algorithm that integrates motion with improved appearance feature information.The research conducted in this paper mainly involves detection-based multiobject tracking algorithms.To address the problem of inaccurate prediction values caused by camera motion in the uniform motion model,this paper uses LK optical flow to track Fast feature points’ motion between images.We then use the RANSAC algorithm to estimate the affine transformation matrix between the previous and current frames,and finally compensate for the camera motion’s effects on the Kalman filter prediction values based on the affine transformation matrix,thereby correcting the problem of inaccurate predictions caused by camera motion and achieving higher real-time performance.In the Kalman filter update phase,we set the observation noise to be adaptively adjusted according to the confidence scores of the detection results to fit the changing observation noise and obtain more accurate object state estimates.In addition,this paper also analyzes the problem of rapid deformation in vehicle tracking,and appropriately adjusts the motion model parameters according to the vehicle tracking task,using the Mahalanobis distance to adapt to the vehicle tracking task.Experiments are conducted on the MOT dataset and KITTI dataset,respectively,validating the effectiveness of our algorithm for pedestrian and vehicle tracking.To address the issue of insufficient discriminative power of appearance features extracted by current feature extraction models,CSPRes Net is introduced as the main backbone network for feature extraction,and a channel attention mechanism(SE module)is added to construct the feature extraction network.The classification cross-entropy loss function is chosen to train the network,transforming the feature extraction task into a classification task.A specific dimension of appearance features is chosen to better represent the object,and feature management is performed using exponential weighted averaging.In addition,a method of constructing cost matrices by fusing motion and appearance feature information is proposed to further improve the accuracy of object tracking using motion and appearance feature information.The feature extraction network is trained on the Market1501 dataset and its effectiveness for tracking tasks is evaluated on the MOT and KITTI datasets.Experimental results on the MOT and KITTI datasets demonstrate that the multi-object tracking algorithm proposed in this paper,which integrates adaptive kalman filter with appearance features,significantly improves the performance of object tracking in motion scenes.The fusion of certain appearance feature information further enhances the effectiveness of multi-object tracking.